Motion Attribution for Video Generation

Antonio Torralba; Despoina Paschalidou; Jonathan Lorraine; Jun Gao; Laura Leal-Taix\'e; Olga Russakovsky; Sanja Fidler; Xindi Wu

REVIEW 5 major objections 4 minor 3 cited by

The paper claims that motion in generated video can be attributed to specific training clips, and that fine-tuning a text-to-video model on the top ten percent of those clips improves temporal consistency and physical plausibility beyond fu

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 10:44 UTC pith:66SB556I

load-bearing objection Worth reading: a practical, clearly-described motion-specific data-selection method for video diffusion, but the attribution claim itself is only validated end-to-end, not causally. the 5 major comments →

arxiv 2601.08828 v2 pith:66SB556I submitted 2026-01-13 cs.CV cs.AIcs.LGcs.MMcs.RO

Motion Attribution for Video Generation

Xindi Wu , Despoina Paschalidou , Jun Gao , Antonio Torralba , Laura Leal-Taix\'e , Olga Russakovsky , Sanja Fidler , Jonathan Lorraine This is my paper

classification cs.CV cs.AIcs.LGcs.MMcs.RO

keywords motion attributiondata attributionvideo generationdiffusion modelsinfluence functionsfine-tuning data selectionoptical flow weightingtemporal dynamics

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to answer which training clips determine how a text-to-video model moves. It proposes Motive, a gradient-based attribution method that computes a motion-specific influence score by masking the diffusion loss with optical-flow magnitudes, so static backgrounds are downweighted. The paper claims that ranking a fine-tuning corpus by this score and keeping the top ten percent improves motion smoothness and dynamic degree compared with random selection or even full-corpus fine-tuning, and that human raters prefer the resulting videos about 74% of the time over the base model. If true, this gives a practical route to curate training data for desired dynamics and to diagnose where motion artifacts come from.

Core claim

The central claim is that motion can be attributed in video generative models by computing gradients of a motion-weighted denoising loss rather than the whole-video loss: when each pixel's loss contribution is weighted by normalized flow magnitude, the resulting gradient inner products rank training clips by their influence on temporal dynamics rather than on appearance. The paper demonstrates that the ranking is useful: fine-tuning a pretrained text-to-video model on the top 10% of clips selected by this ranking raises dynamic degree from about 40-41 to 47.6, matches or exceeds full-dataset fine-tuning, and achieves a 74.1% human preference win rate over the base model. It also reports that

What carries the argument

The central object is the motion-weighted loss mask: optical-flow displacement magnitudes between consecutive frames are min-max normalized to [0,1], bilinearly downsampled to the VAE latent grid, and multiplied into the per-location denoising error before backpropagation; an all-ones mask recovers the ordinary loss. This mask is combined with three efficiency choices: a single fixed denoising timestep with a shared noise draw for all training-query pairs, division by frame count to remove a duration bias, and a random Hadamard projection to 512-dimensional gradient vectors whose cosine similarity is the influence score. The mask is what makes the attribution motion-specific; the single-samp

Load-bearing premise

The whole method rests on the idea that gradients of the original, not-yet-fine-tuned model at one moment of denoising, with one fixed noise pattern, correctly predict which clips will teach motion after fine-tuning; the paper checks the final fine-tuned quality but never verifies that link directly.

What would settle it

Build a synthetic training corpus in which one clip's motion is deterministically reused in the query, then compute Motive scores and check whether that clip ranks first; if it does not, the score is not tracking the source of motion. Alternatively, fine-tune after removing the top 10% of Motive-ranked clips and measure motion smoothness and dynamic degree: no degradation would show the ranking is not causal.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Fine-tuning on the top 10% of attributed clips beats full-dataset fine-tuning on dynamic degree (47.6 vs 42.0) and subject consistency, while preserving motion smoothness.
Motion attribution can filter negative-influence clips: the lowest-ranked clips are mostly static footage, camera-only motion, or cartoon content that dilutes temporal learning.
Because per-clip gradients are computed once and stored as small projected vectors, adding a new query costs one backward pass plus a dot product, making iterative data curation practical after a one-time corpus pass.
Cross-motion overlap patterns are similar across two different large video datasets, suggesting the attribution reflects shared structure in how motion is learned rather than dataset-specific noise.
The same qualitative gains transfer to a larger 5B-parameter model, with dynamic degree rising from 42.0 to 48.3 under Motive-selected fine-tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the ranking is causal, an unstated but direct use is diagnosis: generate a video with a failure, run Motive with that video as query, and the top-ranked training clips are likely sources of the learned bad dynamics; one could remove or downweight them instead of fine-tuning.
The motion weight comes from instantaneous flow magnitude, so the attribution emphasizes current speed over longer-range trajectory coherence; a natural extension is segment-level or event-level attribution that assigns influence to one phase of a motion rather than the whole clip, a limitation the paper itself notes.
Because the method only needs per-example gradients, the same motion-masked influence procedure could be applied to other temporal generative settings, such as world models or audio-video generation, by replacing the flow-based mask with the appropriate motion or energy representation.
The single-timestep choice suggests a testable hypothesis: different denoising timesteps carry different motion information, so rankings computed at very early and very late timesteps may select disjoint clips and could improve fine-tuning in complementary ways.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Worth reading: a practical, clearly-described motion-specific data-selection method for video diffusion, but the attribution claim itself is only validated end-to-end, not causally.

read the letter

You should know this paper is the first practical attempt at motion-specific data attribution for video generation, and as a data-curation tool it seems to work. The authors adapt gradient-based attribution (TracIn/TRAK-style cosine similarity, Fastfood projection, common randomness, ICONS majority voting — all prior work) to video diffusion by adding motion-weighted loss masks from optical flow and a frame-length normalization. The method is clearly described and the ablations are sensible: single timestep, projection dimension, frame-length bias, and a comparison showing they are not just selecting high-motion clips. The downstream gains on VBench dynamic degree (47.6 vs 41.3 for random selection) and the 74.1% human win rate over the base model are real evidence that the selection pipeline improves motion under this evaluation. The limitations section is honest about CFG and segment dilution.

The soft spot is the one the reader flagged. The central claim — that these scores trace generated motion to specific training clips — is never validated against ground truth. No leave-one-out retraining, no counterfactual upweighting, no known-influence benchmark. The ablations only show internal consistency (Spearman vs. full gradients, agreement with multi-timestep). So the selection gains could plausibly come from a correlate, like caption category overlap or motion saliency, rather than the per-clip gradient alignment Motive claims to measure. The evaluation also shares the same motion taxonomy for queries and test prompts, which adds a small circularity. And there are no error bars or repeated seeds on the headline tables; the 47.6 vs 41.3 gap could be seed noise. These are addressable, but they are load-bearing for the attribution interpretation.

That said, the paper is honest about what it does: the framing is largely about data selection for fine-tuning, and under that framing it delivers. I would not desk-reject this. It deserves a serious referee who asks for code/data, repeated seeds or confidence intervals, and either a causal validation of the attribution (even a small leave-one-out on a subset) or a clear reframing as a selection heuristic rather than per-clip influence. The paper is for people doing data attribution in generative models or data curation for video; they will get value from it. I would send it out.

Referee Report

5 major / 4 minor

Summary. The paper introduces Motive, a gradient-based data attribution framework for video diffusion models that aims to attribute generated motion to specific training clips. The method combines motion-weighted loss masks (using AllTracker optical flow), a fixed-timestep/shared-noise single-sample gradient estimator, a Fastfood random projection to D'=512, frame-length normalization, and majority-vote aggregation across query videos to select influential fine-tuning subsets. The authors evaluate Motive by fine-tuning Wan2.1-T2V-1.3B on the top 10% of selected clips from VIDGEN-1M and 4DNeX-10M and report improved VBench dynamic degree (47.6 vs. 41.3 for random selection) and a 74.1% human preference win rate over the base model. Additional results on Wan2.2-TI2V-5B and ablations on projection dimension and frame-length normalization are provided.

Significance. If the central attribution claim holds, this would be the first motion-specific data attribution framework for video generative models, with practical value for data curation and artifact diagnosis. The paper has notable strengths: it is the first to adapt influence-style attribution to temporal dynamics via motion masking; it provides detailed computational efficiency analysis and reproducible hyperparameters; it includes cross-dataset consistency analysis (Fig. 7) and ablations of projection dimension and frame-length normalization; and it demonstrates that the proposed selection pipeline improves both automated and human-judged motion quality. The empirical gains are substantial but, as detailed below, the evidence does not yet establish that the per-clip influence scores themselves are causally predictive of fine-tuning outcomes, rather than merely being a useful selection heuristic.

major comments (5)

[§3.1, §4.2, Eq. (17), Algorithm 1] The paper's central claim is that Motive scores predict which training clips will change motion after fine-tuning. However, the only evidence is the end-to-end selection experiment (Table 1, Table 2): fine-tuning on the top-10% of clips ranked by Motive improves VBench dynamic degree and human preference. This validates a selection pipeline, not the per-clip attribution itself. No leave-one-out retraining, counterfactual upweighting, or known-influence benchmark is provided. Since the selected subset may correlate with caption quality, motion saliency, or text-category overlap with the query set, the causal link between the gradient-alignment score in Eq. (17) and downstream motion is not demonstrated. The 'predictivity' desideratum in §3.1 is asserted but not directly tested.
[Tables 1 and 5; §4.3] Tables 1 and 5 report VBench scores without error bars, confidence intervals, or repeated seeds. The headline dynamic degree gap (47.6 vs. 41.3) comes from a single fine-tuning run per method; given the known sensitivity of video fine-tuning to initialization and data order, this gap could be within seed noise. Table 2 reports human preference percentages but no inter-annotator agreement (e.g., Cohen's kappa) or confidence intervals. The authors should provide variance estimates or multiple seeds to support the quantitative claims.
[§4.4, App. F.1] The ablations validate internal consistency rather than ground-truth agreement. The single-timestep ablation reports ρ=66% agreement with a 10-timestep reference, but that reference is itself a gradient-similarity estimator, not a causal ground truth. The projection-dimension ablation measures agreement with unprojected scores (ρ=74.7% at D'=512), which again is internal. Furthermore, key hyperparameters (tfix=751, D'=512, top-10% budget, majority-vote threshold) are fixed without sensitivity analysis. A reader cannot tell whether the reported fine-tuning gains are robust to these choices.
[App. G.2] The appendix explicitly states that the framework does not account for classifier-free guidance (CFG), which is used at inference time in all standard pipelines, and that temporal segments are diluted by whole-clip averaging. Both limitations directly affect the alignment between the training-time attribution objective and the inference-time behavior the paper claims to explain. The paper should either incorporate CFG into the attribution (e.g., by using the CFG-modified gradient) or temper the claim that Motive identifies clips that 'strongly affect motion' in generated videos, since generation uses CFG.
[App. F and §4.1] The query set and the test prompts are drawn from the same ten motion categories (compress, bounce, roll, etc.). This shared taxonomy creates a possible positive bias: the selection is optimized for categories that are then used to evaluate the fine-tuned model. The paper would be stronger with held-out motion categories or with test prompts that are unseen in the query taxonomy, to demonstrate that the improvement generalizes beyond the curated motion vocabulary.

minor comments (4)

[Table 1] The aesthetic quality column for the base model reads '45.365.7' (missing separator), which should be '45.3 65.7'. Please proofread all tables for formatting.
[App. F.1] The paper states t=751 is 'the midpoint of the 1000-step denoising trajectory' but also says it corresponds to 'scheduler index 500'. Clarify the relationship between the timestep index and the scheduler step, since 'midpoint' and 'index 500' are not obviously the same in flow matching.
[Fig. 6] The x-axis label 'All' appears to duplicate; likely the left figure should be 'Number of Samples' with a legend, and the right figure 'Motion Bin'. Please make the figure caption self-contained.
[§4.4] The frame-length normalization ablation reports ρ=78.0% correlation with video length before standardization and a 54.0% reduction after. It would be helpful to state the post-standardization correlation value explicitly for completeness.

Circularity Check

0 steps flagged

No significant circularity: Motive's scores are defined by gradient cosine similarity, and the fine-tuning outcomes are measured independently; the only self-citation (ICONS for majority voting) is not load-bearing.

full rationale

The core attribution score (Eq. 17) is defined as a cosine similarity between projected, normalized gradients of a motion-weighted loss (Eqs. 15–16). The downstream improvements from fine-tuning on top-scoring clips are measured independently on VBench (Table 1) and via human preference (Table 2); these outcomes do not enter the score definition. Hyperparameters such as tfix=751 and D'=512 are selected by internal consistency with multi-timestep or full-gradient references (§4.4, Fig. 4), not by fitting the VBench numbers, so this is not a fitted-input-called-prediction case. The only self-citation that appears in the method is the majority-voting aggregation from ICONS [Wu et al., 2024] (Eq. 18); it is a simple, independently justifiable aggregation choice and is not load-bearing for the central attribution claim. Appendix G.2 explicitly acknowledges that CFG is not accounted for and that segment-level motion information can be diluted; these weaken the causal interpretation of the scores but do not make the derivation circular. There is no step in which the prediction is equivalent to the input by construction, no uniqueness theorem imported from the authors, and no ansatz smuggled in via self-citation. Score 2 reflects only the minor, non-load-bearing self-citation; the central derivation is self-contained and empirically tested.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

Motive introduces no new physical or conceptual entity; it combines existing gradient-based attribution machinery with motion masks. The main epistemic burden rests on the transfer assumption that initialization-time gradients predict fine-tuning gains, and on the optical-flow masks faithfully isolating motion. Several hyperparameters (tfix, D', top-10%, frame standardization) are chosen by hand or by ablation and are not derived from theory.

free parameters (6)

tfix (single timestep) = 751 (scheduler index 500)
Chosen in §4.4 via agreement with a 10-timestep estimate; affects all influence scores.
D' (projection dimension) = 512
Selected in Fig. 4 as a trade-off between Spearman correlation (≈74.7%) and computational cost.
Top-K percentile / voting threshold = 10% of dataset
Used for fine-tuning budget and majority-vote cutoff in §3.5; affects all downstream results.
Frame standardization = 81 frames at 16 fps
Standardizes clips before attribution and interacts with the 1/F frame-length correction in §3.3.
Fine-tuning learning rate and epochs = 1e-5, 1 epoch (dataset repeated 50x)
Training hyperparameters from DiffSynth defaults; affect the magnitude of improvements.
zeta (numerical stabilizer) = 1e-6
Small constant in Eq. 13 to prevent division by zero; minor but technically a hand-set parameter.

axioms (6)

domain assumption Gradient cosine similarity approximates true data influence (identity-preconditioned influence function)
Inherited from TracIn/TRAK in §3.2; no Hessian correction is computed.
domain assumption A single shared (t, epsilon) preserves relative influence rankings
Introduced in §3.2/Eq. 7; ablation shows 66% agreement with a 10-timestep estimate, not with ground-truth influence.
domain assumption Optical-flow magnitude masks isolate temporal dynamics from static appearance
Core of §3.4/Eqs. 13-16; depends on AllTracker accuracy and min-max normalization, and camera-only motion can dominate.
domain assumption Initialization-time gradients predict which clips improve motion after fine-tuning
The key transfer assumption in §3.5/Algorithm 1; tested only via downstream VBench and human preference, never directly.
domain assumption VBench dynamic degree/motion smoothness and human preference measure motion quality
Used as evaluation in §4.1; custom prompts deviate from standard VBench prompts, and automated metrics may not capture perceptual motion quality.
domain assumption Frame-length normalization 1/F removes duration bias
Proposed in §3.3; supported only by correlation analysis, not a formal derivation.

pith-pipeline@v1.3.0-alltime-deepseek · 22594 in / 11198 out tokens · 110431 ms · 2026-08-03T10:44:47.966753+00:00 · methodology

0 comments

read the original abstract

Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. We present Motive (MOTIon attribution for Video gEneration), a motion-centric, gradient-based data attribution framework that scales to modern, large, high-quality video datasets and models. We use this to study which fine-tuning clips improve or degrade temporal dynamics. Motive isolates temporal dynamics from static appearance via motion-weighted loss masks, yielding efficient and scalable motion-specific influence computation. On text-to-video models, Motive identifies clips that strongly affect motion and guides data curation that improves temporal consistency and physical plausibility. With Motive-selected high-influence data, our method improves both motion smoothness and dynamic degree on VBench, achieving a 74.1% human preference win rate compared with the pretrained base model. To our knowledge, this is the first framework to attribute motion rather than visual appearance in video generative models and to use it to curate fine-tuning data.

Figures

Figures reproduced from arXiv: 2601.08828 by Antonio Torralba, Despoina Paschalidou, Jonathan Lorraine, Jun Gao, Laura Leal-Taix\'e, Olga Russakovsky, Sanja Fidler, Xindi Wu.

**Figure 1.** Figure 1: Motive. Top. Motion-gradient computation (§3.4) has three steps: (1) detect motion with AllTracker; (2) compute motion-magnitude patches; (3) apply loss-space motion masks to focus gradients on dynamic regions. Bottom. Our method (§3.2) is made scalable via a single-sample variant with common randomness and a projection, computed for each pair of training and query data, aggregated (§3.5) for a final ranki… view at source ↗

**Figure 2.** Figure 2: Motion attribution examples. Top: Query clips showing float (left) and roll (right) motions. Middle: Top-ranked positive training samples identified by Motive with high influence scores. Bottom: Negative influence samples with minimal, camera-only motion, or cartoon-style content that conflict with target motions. motion semantics); and Ours w/o motion masking (influence of the entire video level without m… view at source ↗

**Figure 3.** Figure 3: Qualitative Comparisons. We compare four motion scenarios (compress, spin, slide, free fall) across the base model, random selection, and our method. Our approach yields more realistic motion dynamics. Supplementary videos are included. planetary rotation). Negatives are mostly static footage, camera-only motion, or cartoons whose simplified kinematics do not transfer. Our method promotes useful motions an… view at source ↗

**Figure 5.** Figure 5: Impact of Frame-Length Normalization on Motion Attribution. Comparison of top-ranked samples for floating motion query. Left: With proper frame-length normalization, top samples consistently exhibit floating motion (waves, floating objects, surfing). Right: Without normalization, rankings are biased by video length, resulting in no coherent patterns among top samples. 128 256 512 1024 2048 45 55 65 75 Spea… view at source ↗

**Figure 4.** Figure 4: Projection dimension analysis. Spearman correlation between projected and full gradients shows rapid improvement with projection dimension, with 512 providing a strong tradeoff between accuracy and efficiency. Projected Gradients Preserve Influence Rankings. Comparing full gradients for attribution is infeasible at a billion-parameter scale. We reduce dimensionality with structured random projections t… view at source ↗

**Figure 6.** Figure 6: Motive is not simply selecting “motion-rich" clips. Our influence scores are computed via gradients, and training videos are considered influential only when they directly improve the model’s ability to generate the target motion dynamics, not because they contain more motion overall. D. Analysis D.1. Motion Distribution Analysis Motive is not simply selecting “motion-rich" clips: The key distinction is th… view at source ↗

**Figure 7.** Figure 7: Cross-motion influence overlap across datasets. Heatmaps showing the percentage overlap of top-100 influential training samples across motion categories for (a) 4DNEX and (b) VIDGEN datasets. Each cell (𝑖, 𝑗) shows the percentage of motion category 𝑖’s influential data (aggregated from 5 queries per category) that also appears in motion category 𝑗’s top-100 influential samples. The asymmetric nature of the… view at source ↗

**Figure 8.** Figure 8: Illustration of motion query set. We generate near-realistic video queries with Veo-3 across ten motion categories. Each category contains five query videos synthesized with controlled prompts and manually screened for clarity and physical plausibility. Component Complexity Runtime Notes Gradient computation 𝒪(𝐵) per sample Query: ∼ 54 seconds Training: ∼ 150 hours 1 A100 GPU; Single forward+backward pass;… view at source ↗

**Figure 9.** Figure 9: Motion overlay visualization. Comparison of original frames and motion overlays for seven video samples across three time points (early, middle, late). The motion overlay demonstrates the spatial weighting of our motion loss: dynamic regions remain visible, while static backgrounds are attenuated to neutral gray. Takeaway: This provides heuristic intuition into what information our motion attribution focus… view at source ↗

discussion (0)

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation
cs.CV 2026-05 unverdicted novelty 6.0

LaMo adds self-supervised latent motion priors via a motion drift loss during training and motion prior guidance during sampling to boost physical fidelity in video diffusion models like CogVideoX.
Variance Reduction for Expectations with Diffusion Teachers
cs.LG 2026-05 unverdicted novelty 6.0

CARV amortizes upstream diffusion teacher costs over noise resamples with timestep importance sampling and stratified-inverse-CDF sampling, delivering 2-3x effective compute gains in text-to-3D experiments and order-o...
Variance Reduction for Expectations with Diffusion Teachers
cs.LG 2026-05 unverdicted novelty 5.0

CARV introduces a hierarchical Monte Carlo estimator with amortized reuse, importance sampling, and stratification that yields 2-3x effective compute gains on diffusion-teacher pipelines while cutting gradient varianc...

Reference graph

Works this paper leans on

4 extracted references · cited by 2 Pith papers

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...
[3]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...
[4]

# o ^ t P ^5 = E

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

arXiv 2022

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

[3] [3]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

[4] [4]

# o ^ t P ^5 = E

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

arXiv 2022